Research On Methods Of Specific People Searching In The Surveillance Video

Specific person searching has always been a popular and also a challenging technology in the field of computer version. This technology, which covers computer version, video processing and image segmentation, related to much technology of intelligent video surveillance and video retrieval technology. In surveillance videos, specific person searching requires the computer to recognize a person with different stances, which is a difficulty of current computer and video processing technology.Combining with the advanced image processing and video processing technology, this paper discusses the related technology in specific people searching field based on in-depth study of such systems. These techniques include background modeling, moving target detection, and matching of features. In moving target detection, this paper compares several current algorithms, and a Surendra algorithm based on HSV color space is proposed. In feature matching, we studied global features and local features as well as their applicable situations. Especially the SIFT feature is studied deeply and be experimented.Since we can not take use of the advantages of single features in specific person searching, we propose a method of multi-feature fusion. It considers the global feature and local feature, which maintains the correctness of single-feature, used in content retrieval, but also takes into account the impact of retrieval results of different features. This paper proves that the combined feature is better than single-feature when used in content retrieval.In experiments, this paper implements a specific person searching system containing two sub-models:moving target detection model and feature detection and matching model. The former is used to detect the moving targets in surveillance videos with background updating algorithm. The latter is the key model of this system, which is responsible for detecting the global feature and local feature, and than complete the retrieval task according appropriate matching algorithm. The simulation experimental results show the feasibility of the method we proposed.